To maybe add four point:
(i) I strongly disagree with Lior's claim that "many studies _with_ replicates have shown that for a large range of expression values biological variability is small compared to 'shot noise'". Once a gene has more than 1,000 reads, Poisson noise drops below 3%, and then, biological variation of even only a few percent dominates technical noise. I have so far never seen any set of biological replicates with variation below a few percent, and, barring really old data, all RNA-Seq experiments contain plenty of genes with more than thousand counts.
(ii) Whenever you perform any kind of comparison between samples, you must know the typical amount of biological variation. Ideally, you get this information from comparing replicates, but to a certain extent, past experience may be at least somewhat admissible. Hence, when Simon Andrews mentions above "results which are so startlingly clear there's no conceivable chance that they're just happening through any kind of noise", he is implicitly comparing with the amount of variability he has seen in previous experiments, which, he asserts, is vastly smaller. If one has performed very similar experiments before, one might even be justified to plug in dispersion estimates from previous studies into the analysis of a current one without replicates. However, as this requires the reader to believe you that you performed the new experiment with the same care and luck as the old one, I would not accept such an argument in a publication. For a pilot study, it may be a useful approach, though.
(ii) On the other hand, I consider it perfectly acceptable to sequence only one sample per condition if one then performs additional validation (e.g. by qPCR) on further samples for a sufficiently large panel of representative genes.
(iv) Cost can never be an argument. Either you need replicates or you don't. If you need them and cannot afford them, you have no business doing the experiment.
(i) I strongly disagree with Lior's claim that "many studies _with_ replicates have shown that for a large range of expression values biological variability is small compared to 'shot noise'". Once a gene has more than 1,000 reads, Poisson noise drops below 3%, and then, biological variation of even only a few percent dominates technical noise. I have so far never seen any set of biological replicates with variation below a few percent, and, barring really old data, all RNA-Seq experiments contain plenty of genes with more than thousand counts.
(ii) Whenever you perform any kind of comparison between samples, you must know the typical amount of biological variation. Ideally, you get this information from comparing replicates, but to a certain extent, past experience may be at least somewhat admissible. Hence, when Simon Andrews mentions above "results which are so startlingly clear there's no conceivable chance that they're just happening through any kind of noise", he is implicitly comparing with the amount of variability he has seen in previous experiments, which, he asserts, is vastly smaller. If one has performed very similar experiments before, one might even be justified to plug in dispersion estimates from previous studies into the analysis of a current one without replicates. However, as this requires the reader to believe you that you performed the new experiment with the same care and luck as the old one, I would not accept such an argument in a publication. For a pilot study, it may be a useful approach, though.
(ii) On the other hand, I consider it perfectly acceptable to sequence only one sample per condition if one then performs additional validation (e.g. by qPCR) on further samples for a sufficiently large panel of representative genes.
(iv) Cost can never be an argument. Either you need replicates or you don't. If you need them and cannot afford them, you have no business doing the experiment.
Comment